A Network Coding Approach to Overlay Network Monitoring Christina Fragouli
Athina Markopoulou
´ Ecole Polytechnique F´ed´erale de Lausanne
[email protected]
Stanford University
[email protected]
September 28, 2005 Abstract Monitoring and diagnosis of network conditions is a central problem in networking. As such, it has received a lot of attention in the Internet community in general and in the context of overlay networks in particular. Independently, recent advances in network coding have shown that it is possible to increase network capacity and better share the available resources by allowing intermediate nodes to perform processing operations, in addition to just forwarding packets. In this work, we propose the use of network coding techniques to improve several aspects of network monitoring in overlay networks. As a specific application, we use our approach for the well-known problem of network tomography, and in particular for inferring link loss rates from end-to-end measurements. We demonstrate that our approach can decrease the bandwidth used by probes, improve the accuracy of estimation, and decrease the complexity of selecting paths or trees to send probes.
1
Introduction
Distributed Internet applications often use overlay networks, that enable them to detect and recover from failures or degraded performance of the underlying Internet infrastructure. To achieve this high-level goal, it is necessary for the nodes in the overlay to monitor Internet paths, assess and predict their behavior, and eventually make efficient use of them. Clearly, accurate monitoring at minimum overhead and complexity is of crucial importance for the operation of all networks, and particularly for overlay networks [1]. In the past decade, several approaches have been proposed for inferring network characteristics of interest (such as topology, packet loss rate, delay, and failures) using end-to-end measurements [2, 3, 4]; this class of problems is commonly referred to as network tomography. Active measurement techniques have been proposed that send sequences of probe packets from a set of sources to a set of receivers, and infer link-level metrics of interest from the received packets. Some techniques send probes over unicast paths [5] while others use multicast trees [2, 3]; to cover the entire network, a mesh of paths and/or trees is needed. The bandwidth efficiency of these methods can be measured by the number of probe packets needed to estimate the metric of interest within a desired accuracy. It depends both on (i) the choice of paths/trees over which sequences of probes are sent and on (ii) the number of probes in each sequence. Clearly, there is a tradeoff between bandwidth efficiency and estimation accuracy; it is desirable to improve both as well as to keep computational complexity low. In this work, we propose the use of network coding techniques [6, 7] to improve several aspects of network monitoring. The basic idea of network coding is to allow intermediate nodes to process the incoming packets before forwarding them. The set of operations that intermediate nodes perform are referred to as a network code; typically, linear codes are used [7]. The idea of network coding (albeit difficult to apply to today’s Internet routers) can be gracefully applied to overlay networks, where the network designer has control over the intermediate nodes in the overlay; furthermore, we envision the use of network coding only for special probe packets and not for forwarding regular traffic.
Allowing nodes in an overlay network to perform network coding can improve all aspects of network tomography, namely bandwidth usage, estimation accuracy, and complexity in choosing which paths to monitor. More specifically, the use of network coding allows to (i) eliminate the overlap between paths and/or trees needed to cover the entire network (ii) use less probes per sequence to achieve a certain accuracy, by intelligently using not only the number, but also the content of received probes and (iii) reduce the complexity of choosing which paths to monitor. In general, we believe that the idea of combining network coding with network tomography techniques is very promising. As a concrete example, we show how to use simple linear coding to improve the inference of link-level loss rate from end-to-end measurements. The structure of the paper is as follows. Section 2 summarizes related work in monitoring and network coding. Section 3 discusses a motivating example, which demonstrates the key points of our approach. Section 4 discusses more formally the application of network coding to network tomography. Section 5 presents simulation results and section 6 concludes the paper.
2
Related Work
Within the broad area of network monitoring, we are interested in network tomography for overlay monitoring, i.e., in using active (unicast or multicast) probes between overlay nodes to infer characteristics of the Internet paths between them. There has been a significant amount of work in this area. Our novel contribution is the application of ideas from network coding. In the context of resilient overlay networks (RON) proposed in [1], O(n2 ) paths are monitored, where n is the number of end-hosts. The authors in [5] chose a basis of k
